Facemask Wearing Correctness Detection Using Deep Learning Approaches

Abstract

The occurrence of widespread fatality, inadequate medical infrastructure and the infectious nature of the COVID-19 virus have necessitated the formulation of appropriate risk minimization methods including extensive use of technology. In the absence of effective antiviral and limited medical resources, among the measures recommended by the World Health Organization (WHO), wearing a mask is considered to be an effective non-pharmaceutical intervention that can be used to prevent the spread of the COVID-19 virus. Hence proper wearing of the mask and its effective monitoring in public placed for reliable enforcing of the community level protocol, adoption of technology becomes crucial. To contribute towards communal health, this paper aims to report the design of an accurate and real-time technique that can efficiently detect non-mask faces in public and thus, suggest ways to formulate measures for enforcing proper wearing of the mask. Among many such technologies, internet of things (IoT) and artificial intelligence (AI) have emerged as viable ones as these combine a host of sensor packs, wireless communication based networking and automated decision making. Further emerging applications involving edge computing, deep learning (DL) and Deep Transfer Learning (DTL) have enabled IoT to take part in a decisive role in the health care sector and help to minimizing damages related to pandemic situations. The presented framework is based on Artificial Neural Network (ANN) tools that use hand–crafted feature samples and DL techniques like Convolutional Neural Network (CNN) and a specialized CNN called YOLO and Support Vector Machine (SVM) classifiers. Here MobileNet has been used as a baseline method which is extended by applying the concept of transfer learning to fuse high-level semantic information in multiple feature maps with samples from Real World Masked Face Dataset. In addition, we also propose a bounding box transformation to improve localization performance during mask detection. It is observed that the proposed technique achieves high accuracy (97.2%) when implemented with MobileNet.